Document resource
Background High Resolution Manometry (HRM) is the gold standard for diagnosing esophageal motility disorders, but its interpretation is time-consuming (typically >10 minutes per study) and suffers from substantial inter-observer variability, especially for achalasia subtypes. This study developed a convolutional neural network (CNN) to automatically classify achalasia (AC) subtypes 1–3, esophagogastric junction outflow obstruction (EGJOO), and normal subjects from single-swallow HRM images, aiming to provide a rapid and accurate computer-aided diagnostic tool.Methods We extracted representative single-swallow HRM images from 563 consecutive studies (Medtronic device, ManoView 3.0.1 software) and categorized them into five classes (normal, EGJOO, AC type 1/2/3) according to the Chicago Classification v4.0. Using an EfficientNet-B0 backbone, images were split into training (n=365), validation (n=85), and test (n=113) sets (8:2 ratio). Training lasted 50 epochs (batch size=8, initial LR=1e-4) with a cosine annealing scheduler. Data augmentation included random flipping, rotation, and brightness-contrast adjustments. The final model had 4.41 million parameters. Performance was evaluated using accuracy, precision, F1-score, AUROC, and inference time ( IDDF2026-ABS-0539 Figure 1).Results The model achieved 90.27% (F1=90.41%) overall accuracy (IDDF2026-ABS-0539Figure 2(A,G)), with high confidence in correct classifications (IDDF2026-ABS-0539 Figure 2(B,D,F)). The macro-average AUROC was 0.9965 (IDDF2026-ABS-0539 Figure 2(C)), with near-perfect performance for EGJOO (AUROC=1.000) and AC3 (0.998), while AC2 was relatively lower (F1=0.83, AUROC=0.989). Training showed stable convergence with no overfitting (IDDF2026-ABS-0539 Figure 2(E)). Misclassification primarily involved AC1 vs AC2 (45.5% of errors; IDDF2026-ABS-0539 Figure 2(F)), with an overall error rate of 9.7%. Inference time was 45 ms for a single image (IDDF2026-ABS-0539 Figure 2(H)), supporting real-time clinical application.Conclusions We successfully developed a deep learning model using EfficientNet-B0 for automatic five-class classification of HRM images from a standardized clinical platform (Medtronic, ManoView 3.0.1). With near-instant inference (45 ms/image) and high diagnostic accuracy, this model holds promise as a real-time assistive tool for esophageal motility disorders, potentially reducing expert workload and improving diagnostic consistency. External validation in multi-center cohorts is warranted before clinical deployment.Abstract IDDF2026-ABS-0539 Figure 1Abstract IDDF2026-ABS-0539 Figure 2